Slashdot Mirror
GDDR2 Emerging As A Real Standard
An anonymous reader writes "I noticed here that EE Times is reporting that the GDDR2 standard is finally becoming a reality. Both NVIDIA and ATI's latest chips offer support. ATI helped spearhead the initiative to develop the standard. The significance of this is great, since it may very well mean that every 18 months or so a new graphics memory standard will be released."
But it seems like this whole 'building names on each other' thing is getting out of hand.
GDDR2 SDRAM? What the hell is that supposed to mean? Sheesh. Why can't you just call it something like DDR3 or GDRAM or something simple like that?
using namespace slashdot;
troll::post();
Comment removed based on user account deletion
"GDDR3 will consume half the power of GDDR2 and operate up to 50 percent faster."
Maybe graphic makers should hold out on GDDR2 for GDDR3. People that buy high-end graphics cards want quality. Look at the GeForce FX. It's going to kill NVIDIA. I think NVIDIA and others (ATI) are going to really learn from the FX and make extra sure that what they come out with will be real innovation, not a quick way to get back on top and the expense of their customers.
It would be nice if that EEtimes article even gave a slight, non-indepth, technical description of what exactly GDDR2 is.
Can anyone answer me that? What makes it special?
SuPz.orG
Isn't it interresting how graphics adapters use the fastest memory available these days, not the CPU. Not counting L1/2/3 caches that is...
This really goes to show how humans are visual animals above all. I wonder how much more power could be squeesed out of porcessors if we were to use memory like this and wider buses...
.: Max Romantschuk
The good news in the article is that the much "better" memory GDDR3 will be standardized from the beginning with may suppliers and hopefully a lower price. Forget GDDR2!
I think they only say that because GDDR3 is farther off into the future.
I've noticed once these things get closer to an actual release date, these people tend to take off their rose colored glasses. My money says there won't be much of a difference between the two different memory types when they're actually released. Not enough to justify what will most definitely be a much higher price for GDDR3.
if you are going to do that, you should also factor in the grief of trying to install a bleeding edge card with bleeding edge drivers.
Having wasted a lot of time and multiple re-installs, I now stick to "not quite bleeding, but still a bit bloodstained" edge products, where at least the drivers are mature.
Humorous signatures are over-rated.
Actually there is a difference in the way CPU and GPU see
memory.
A CPU cares a lot about latency because typical code will
have "random" accesses scattered with calculations in
between. The same data and code areas are often
accessed many times and data are small
(e.g. a Word document is small) while code
maybe quite large.
That's why CPU's don't have enormous
256-bit buses (which have the same latency as a 64-bit
bus)
A GPU performs "multimedia" calculations which typically
involve serial access to memory where caching can be of
very little help. You cannot "cache" a whole texture set
and code is of really trivial size (until now, maybe
PixelShader 2.0+++ will change all that). Therefore
a GPU needs serial access to huge areas of memory,
involving items of similar size and in regular intervals.
That's why a GPU needs BANDWIDTH (not necessarily
latency, because when the calculation starts latency
is hidden inside the calculation loop).
Considering the above, P4 is a "multimedia" design (much
more like a GPU) that's why it was made to work with
very high FSB and RAMBUS (high bandwidth) originally.
Contrary to this, AMD Athlon is a "generic" design which
does not depend on huge bandwidth but on very low
latency (hence the HUGE L1 cache). That's why P4 needs
HyperThreading : its long pipelines do not care a lot about
latency but can cause a big bottleneck if they stall. Intel
feeds them continuously by drawing instructions from 2
processes at once (so that the pipeline does not remain
empty if one process is stalled from the front side bus or
something...).
Anyway, I expect GPUs to drift slowly towards the generic
CPU design because pixelshader language has become
quite complicated with long loops etc. Gradually this
means that GPUs (esp. with DirectX9) will start being
compute-limited and not texture-fill-rate limited
(anything over 2 GTexel/s is really absurd for
typical screen sizes). This will propably become apparent
with DOOM III.
P.